Diffusion processing

This page discusses the preprocessing steps required to prepare a diffusion-weighted dataset for analysis using TractoR, as well as subsequent operations that can be performed to manipulate diffusion gradient directions and fit models to the data.

The pipeline

The preprocessing steps required to run neighbourhood tractography or other analysis on a diffusion-weighted data set using TractoR are broadly the same as those required by the FSL diffusion toolbox (FDT) for tractography. It is therefore quite possible to perform all of these steps independently of TractoR, and the user may decide to use the relevant FSL tools, or their equivalents from another package. However, reasons to use TractoR might include the following:

Whichever method is used to perform them, the requisite steps are as follows.

  1. Convert DICOM files from an MR scanner into a 4D data set file in Analyze or NIfTI format. The result is an image called rawdata (with appropriate suffix depending on the file type) in the diffusion subdirectory of the TractoR session hierarchy. The file directions.txt, which describes the diffusion weighting applied to the images, will also be created if possible. If the relevant information is not available, this latter file must be created manually—if this is the case then TractoR will produce a warning.
  2. Optionally correct for susceptibility-induced distortions with FSL topup, then identify an image volume with little or no diffusion weighting, to be used as an anatomical reference. By default this file will be called refb0 and stored in the diffusion subdirectory.
  3. Create a mask which covers only that part of the refb0 volume which is within the brain. Skull and other nonbrain tissue is left outside this mask. The image file mask is created in the diffusion subdirectory, as well as a masked version of the anatomical reference image, called maskedb0.
  4. Correct for eddy current-induced distortion effects in the data, using the anatomical reference volume as a registration target. This step currently uses one of the FSL tools eddy or eddy_correct, or the internal NiftyReg interface, and produces a file called data in the diffusion subdirectory.

In addition to these steps, some kind of diffusion model fitting is a prerequisite for most subsequent analysis. Steps for performing such fits, and other common data manipulations, are laid out below.

Sorting DICOM files

A single scan session often comprises a range of image acquisitions of different types, for example including structural and functional imaging as well as a diffusion-weighted sequence. It is therefore a common first step to sort the DICOM files corresponding to the diffusion scan out from everything else. Assuming for present purposes that data sets are stored as subdirectories of /data, and that each data set has its own dicom subdirectory containing all associated DICOM files, this sorting process can be achieved by typing something like

cd /data/subject1/dicom
tractor dicomsort

This will result in various new subdirectories being created, labelled with their appropriate series number and description. It is up to the user to identify which series corresponds to the diffusion data.

Using the dpreproc script

With DICOM sorting already performed if necessary, running the preprocessing pipeline for a single session directory is a matter of typing something like

cd /data/subject1
tractor dpreproc

By default, TractoR will assume that all DICOM files it finds under the main session directory, /data/subject1, relate to your DTI acquisition. If in fact your diffusion DICOM files are stored in some subdirectory, perhaps /data/subject1/dicom/dti, you should tell TractoR this by instead using

tractor dpreproc DicomDirectories:dicom/dti

Notice that the DICOM subdirectory given is relative to the session directory. More than one directory can be specified if multiple acquisition series are relevant.

Alternatively, the newer divest back-end can be used to select the relevant series interactively, using

tractor dpreproc DicomReader:divest

The preprocessing can be completed noninteractively by setting the “Interactive” option to false:

tractor dpreproc Interactive:false

(Note, however, that in this case default parameters will be used, without the chance to check that the results are appropriate. Some parameters can, however, be adjusted using other options to the dpreproc script: run tractor -o dpreproc for details.) By default, TractoR will run all four stages, but will miss out any stage that has already been successfully completed. To run every stage except the final one, type

tractor dpreproc RunStages:1-3

or to start from the beginning again even if some stages have already been done, use

tractor dpreproc Force:true

If you want to find out which stages have already been run, simply give

tractor dpreproc StatusOnly:true

Checking and rotating gradient directions

The directions of applied diffusion-weighting gradients are determined from the DICOM files if possible, during stage 1 of the dpreproc script. However, if dpreproc is not used, or the gradient directions cannot be found, it may be necessary to specify them manually. To do so, the directions should be put into a plain text file, arranged either one-per-column or one-per-row, normalised or unnormalised, and with or without zeroes for b=0 measurements. The gradread script can then be called, passing the session directory, gradients file and the big and little b-values:

tractor gradread /data/subject1 /data/directions.txt 1000 0

This should normally only be necessary once for a given sequence, since a cache is automatically used to store gradient directions for use with other data sets acquired the same way.

Whichever way the gradients are initially obtained, it is a good idea to check that the signs of the directions are correct. The gradcheck script assists with this by showing the principal directions of diffusion in three representative slices of the brain, for checking against the user’s expectations. It can be run from the session directory with

tractor gradcheck

Another step which is commonly performed is gradient rotation, to compensate for the registration performed to correct for eddy current effects in dpreproc stage 4. If required, this should be the last step performed on the directions. It is run with simply

tractor gradrotate

The plotcorrections script can be used to see how big the effect of this step will be: run tractor -o plotcorrections for more information.

Model fitting

Fitting diffusion tensors is a standard processing step for diffusion-weighted data, and results in the creation of a range of derivative images, including maps of fractional anisotropy and mean diffusivity. This fitting can be performed using

tractor tensorfit

There are four alternative approaches to fitting the tensors available, but standard least-squares fitting is the default: see tractor -o tensorfit for details. The Camino toolkit offers many more methods.

The diffusion tensor is a very limited model, however, particularly for tractography. It is therefore recommended to run FSL BEDPOSTX algorithm to fit a “ball-and-sticks” model and generate samples for probabilistic tractography. This typically takes several hours. The command for running this is

tractor bedpost

The underlying FSL bedpostx program takes a parameter which determines the maximum number of fibre populations which may be represented for each voxel. The larger this value, the longer bedpostx will take to run, but if set to 1 then no crossing fibre information will be available. The default value is 3, and this can be changed using the NumberOfFibres option:

tractor bedpost NumberOfFibres:2

It is important to note that the number of fibres fitted is a property of the session, and so once set it cannot be changed without processing the data again. If you wish to try different values of this option on a single data set, you will need to duplicate the session hierarchy, since these would be considered two distinct preprocessing pipelines, producing two different data sets.

The status script

To find out information about a particular session directory and the data stored within it, you can use the status script, which produces output like the following:

  Session directory        : /usr/local/tractor/tests/data/session
  Working directory exists : TRUE

  Preprocessing complete        : TRUE
  Data dimensions               : 96 x 96 x 60 x 12 voxels
  Voxel dimensions              : 2.5 x 2.5 x 2.5 mm x 1 s
  Number of shells              : 1
  Diffusion b-values            : 0, 1000 s/mm^2
  Number of gradient directions : 2, 10
  Fibre orientation model(s)    : DTI, FSL-BEDPOSTX (1 fibre per voxel)